The harness landscape

The shared languages of the harness: how agents discover tools (MCP), talk to each other (A2A, ACP) and drive user interfaces (AG-UI). Protocols are what keep the layer from fragmenting into silos.

The core of the stack: frameworks and SDKs that structure the agent loop — plans, tools, state, handoffs, guardrails and control flow. From graph-based runtimes (LangGraph) to Big Tech SDKs (OpenAI Agents, ADK, Strands) and minimal cores (smolagents, PocketFlow).

Everything past the context window: persistent memory, knowledge graphs and context engineering that let agents remember across sessions and ground decisions in real data.

Ingestion, crawling, search and retrieval — feeding agents grounded context: RAG frameworks, web crawlers and search APIs purpose-built for LLM consumption.

The hands of the agent: browser automation, tool registries, integration platforms and MCP servers that connect agents to the outside world.

Where agent code actually executes: secure sandboxes, code-interpreter clouds and serverless agent runtimes that make tool calls safe and fast.

The measurement half of harness engineering: tracing, testing, red-teaming and continuous evaluation of agent behavior in dev and production.

Guardrails, permissions and red-teaming — keeping powerful, tool-using agents inside the lines their operators drew.

The substrate below the harness: durable execution, queues and workflows that survive crashes, plus gateways that route and proxy model traffic.

The flagship vertical: agents that write, review and ship software — from autonomous SWE agents to IDE companions and terminal-first harnesses.

Harnesses pointed at a domain: customer experience, finance, back-office work and the 'AI workforce' platforms shipping agents to end customers.

The meetups, conferences and hackathons where harness engineering know-how spreads — tracked in the Events directory.